d1-3B
A decision model released by Liquid AI and part of the same d1 series as d1-omni-600M. It has approximately 3B parameters and achieves a high score on the Decision Index.
Timeline (last 90 days)
When selecting a model, it is recommended to first run a baseline with a rule engine; if the rules can cover 80% of cases, the remaining long tail can be handed to d1-3B.
The positioning of d1-3B is not to pursue comprehensive general-purpose dialogue, but to complete high-frequency small decision tasks such as classification, routing, and judgment with extremely low latency on low-power hardware such as laptops.
The task types for which d1-3B is suitable include intent classification, tool routing, content gating, and state judgment.
On Jetson AGX Thor, d1-3B takes 16 milliseconds to answer one question and 20 milliseconds for three questions; on Jetson AGX Orin 64GB it is 26 milliseconds and 35 milliseconds; on Jetson Orin Nano it is 50 milliseconds and 73 milliseconds.
The d1-3B model has about 3.1 billion parameters, is trained on the LFM2.5-VL-3B vision-language backbone, and supports text and image input.
According to data Liquid provided in the Hugging Face release post, d1-3B takes 8 milliseconds to answer one question and 21 milliseconds for three questions on an NVIDIA RTX 4090.
d1-3B scores an average of 82.9 on seven public text benchmarks and 48.57 on the public split of Decision Index v0.2.1; these are vendor data and need to be verified by independent evaluation.
在 Liquid AI 的 Decision Index 0.2.1 上，d1-omni-600M 得分 15.95，d1-3B 得分 48.57，Decider 2B 得分 28.97。
Liquid AI 于 10 月 7 日发布 d1-omni-600M 与更大的 d1-3B，官方将其定位为实验性检查点。
Liquid 官方表示不报告 d1-omni-600M 的推理速度数据，博客中提到的亚 50ms 延迟仅适用于 d1-3B。
11 more timeline events
All Facts (20)
d1-3B's question types include noul (returns P(yes)), choice (returns the selected option, confidence, and the probability of each option), and score (returns the expected rating, confidence, and probability distribution on a 2–10 ordinal scale).
60%Unverifiedd1-3B on text tasks: SQuAD 2.0 85.3, Civil Comments 93.0, MASSIVE intent 87.3, PubMedQA 66.0, BoolQ 86.7, XNLI 85.0, PAWS-X 76.9, averaging 82.9 across the seven.
60%UnverifiedThe task types for which d1-3B is suitable include intent classification, tool routing, content gating, and state judgment.
50%UnverifiedThe positioning of d1-3B is not to pursue comprehensive general-purpose dialogue, but to complete high-frequency small decision tasks such as classification, routing, and judgment with extremely low latency on low-power hardware such as laptops.
50%UnverifiedWhen selecting a model, it is recommended to first run a baseline with a rule engine; if the rules can cover 80% of cases, the remaining long tail can be handed to d1-3B.
50%UnverifiedOn Jetson AGX Thor, d1-3B takes 16 milliseconds to answer one question and 20 milliseconds for three questions; on Jetson AGX Orin 64GB it is 26 milliseconds and 35 milliseconds; on Jetson Orin Nano it is 50 milliseconds and 73 milliseconds.
50%Unverifiedd1-3B scores an average of 82.9 on seven public text benchmarks and 48.57 on the public split of Decision Index v0.2.1; these are vendor data and need to be verified by independent evaluation.
50%UnverifiedThe d1-3B model has about 3.1 billion parameters, is trained on the LFM2.5-VL-3B vision-language backbone, and supports text and image input.
50%UnverifiedAccording to data Liquid provided in the Hugging Face release post, d1-3B takes 8 milliseconds to answer one question and 21 milliseconds for three questions on an NVIDIA RTX 4090.
50%Unverified在 Liquid AI 的 Decision Index 0.2.1 上，d1-omni-600M 得分 15.95，d1-3B 得分 48.57，Decider 2B 得分 28.97。
50%UnverifiedLiquid 官方表示不报告 d1-omni-600M 的推理速度数据，博客中提到的亚 50ms 延迟仅适用于 d1-3B。
50%UnverifiedPublished latency: RTX 4090 (bf16) 8 ms for a single question, 17 ms for a 384px image, and 475 states/s batched throughput; Apple M5 Pro (MPS) 30 ms / 62 ms / 78 states/s.
50%UnverifiedLiquidAI/d1-3B has accumulated 10,843 downloads on Hugging Face since its release.
50%Unverifiedd1-3B is a decision model that reads text, JSON, and images and directly gives answers to predefined questions (yes/no, multiple choice, scale rating) in a single forward pass, generating no text, with an output token count of 0.
50%UnverifiedLiquid AI's Open d1 series includes d1-3B (text + image) and d1-omni-600M (text + image or text + audio, experimental).
50%Unverifiedd1-3B has 3.12B parameters, its vision encoder is a shape-optimized version of SigLIP2 NaFlex (400M), its context length is 32,768 tokens, its vocabulary is 128,000, its weight format is Safetensors/BF16, its base model is LFM2.5-VL-3B (the text backbone was weight-averaged with LFM2.5-2.6B), and its license is lfm1.0.
50%Unverifiedd1-3B scored 48.57 on the Decision Index v0.2.1 (public split), behind Winnow-12B (50.02) and ahead of Decider 35B-A3B (47.11).
50%Unverifiedd1-3B averages 74.1 across 11 image benchmarks, versus 73.9 for LFM2.5-VL-3B; on VL-RewardBench d1-3B scores 65.0, far ahead of LFM2.5-VL-3B's 50.9, but it lags on CV-Bench (82.1 vs 87.6), VisualWebBench (71.4 vs 78.3), and POPE (88.5 vs 90.1).
50%UnverifiedLiquid has not published a comparison of d1-3B with Qwen2.5-VL-3B or Gemma, nor has it published calibration error data (such as ECE or reliability diagrams).
50%UnverifiedLiquid AI 于 10 月 7 日发布 d1-omni-600M 与更大的 d1-3B，官方将其定位为实验性检查点。
50%